Rozobox: resilient distributed storage on 8 RPis

RozoBox is a cluster of 8 Raspberry Pis (RPis) used to experiment, teach and popularize distributed storage. It was part of a project driven by two students from Polytech Nantes: Olivier Blin and Quentin Lebourgeois, and supervised by Nicolas Normand, and myself.

RozoFS was deployed on the RPis and leveraged their 8 SD cards to provide a single storage volume that one can mount through the network.

To be fault-tolerant, RozoFS distributes redundant data across multiple servers. To that end, when a movie is stored on RozoFS, it is split into multiple data blocks. Each block is itself cut into data chunks, and some redundant chunks are computed from them. For instance, to provide fault-tolerance against two failures (like RAID-6), RozoFS computes two redundant chunks from four data chunks (and we get 4+2 chunks). Then, each chunk is spread on different servers. For reading, it is necessary to reach any subset of four chunks among the six that were distributed during the write process.

To make things clearer, the students have developed a monitoring tool based on ncurses which displayed the 8 RPis and enables give a quick statement of each card regarding CPU/RAM/disk/network metrics and according to colors:

1. green: available (through the network) and idle;
2. yellow: available and working (read/write activity);
3. red: unavailable.

Thus, the demo was the following. I mounted the RozoFS volume from my computer after plugging it to the network switch. Then I played a video stored on RozoFS and the monitoring tool showed that 4 cards turned from green to yellow, meaning they were implied in the video playing. Then I caused a failure by disconnecting one of those implied card from the network. The video continued to be played while the monitoring tool showed one reading card turning red, and another one turning from green to yellow, meaning a first set of redundant chunks are used.